Automated analysis of the photoreceptor mosaic in VIVO

• Main Author: Turpin, Alan P.
• Published: University of Ulster 2012
• Subjects: 617.73507
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.588497

The retinal photoreceptor mosaic consists of millions of cones and rods that when struck by light produce a chemical reaction which in turn is sent to the brain and is interpreted into what the eye sees. The density of such retinal cones could be indicative of underlying diseases. Visual analysis of the retinal photoreceptor mosaic in vivo can be both subjective and time consuming. Through the use of image processing techniques, the automatic segmentation and analysis of the photoreceptor cones would be beneficial. Within this thesis two approaches are presented for the automatic identification of photoreceptor cones in retinal images in both 2D and 3D images. These algorithms apply a process of multi-scale modelling and normalised cross-correlation to identity where the retinal cones are located. The retinal images can also contain blood vessels and within these vessels blood cells can sometimes be visible. The cone detection algorithms can in some instances identify the blood cells incorrectly as retinal cones and therefore a process of identifying and segmenting the blood vessels is proposed. The manual identification of the photoreceptor cones in retinal images is time consuming; therefore to evaluate the identification algorithms, a process of synthetic data generation is illustrated. This process randomly generates and places the multi- scale models in an image with varying amounts of noise and blurring in order to closely resemble real images. The application of the proposed novel 2D cone detection algorithms to a sample of 10 real images correctly identified 93.21 % of the retinal cones with a false positive rate of only 6.33%. The 3D algorithm correctly identified 82.02% of the retinal cones with a false positive rate of 1.72%. These results illustrate that the performance of the 2D approach for retinal cone identification is comparable to that reported in current research of 97% accuracy and 3% false positives. The 3D approach misses more of the retinal cones, however identifies significantly fewer false positives.